AI can make thinking less linear

The risk is not that AI recombines what already exists. Humans do that too. The risk is using it to reach the average before we have explored enough of the territory.

One of the most persistent arguments against AI is that it cannot be creative because it only works from material it has already encountered.

It learns from existing writing, images and ideas, then predicts a plausible way to put those things together. Ask it to produce something original and it often gives you a polished average: familiar structure, familiar language, familiar conclusion.

All of that is a reasonable description of the limitation.

I am less convinced by the conclusion people draw from it.

Human creativity does not usually arrive from an untouched place either. We remember, imitate, adapt, misremember and recombine. We absorb structures so deeply that we stop noticing their sources. We encounter an old idea in a new setting and experience the connection as original.

This does not make a human brain the same thing as a large language model. A person has a body, experience, intention, emotion, responsibility and a life outside the material being processed. A model has none of those things.

But if the objection is simply that AI builds from what came before, then the standard would disqualify a great deal of human culture too.

The more useful question is not whether the raw material is inherited. It is what gets connected, why the connection matters and who decides what should be made from it.

Originality was never a clean-room process

Quentin Tarantino is an obvious example because his influences are so visible.

The British Film Institute’s account of Reservoir Dogs traces recognisable elements to The Killing, The Taking of Pelham One Two Three and City on Fire, alongside the wider influence of thrillers, westerns, blaxploitation, Hong Kong action films and the French New Wave.

That does not settle every argument about homage, borrowing or plagiarism. It does show why originality cannot mean the absence of precursors. Tarantino’s work is distinctive because of the combination: the references he selects, the genres he collides, the dialogue he writes, the chronology he breaks and the sensibility he applies.

The same is true of stories more broadly. Greek myths continue to supply familiar templates: the journey, the fall, the impossible choice, the family curse, the attempt to escape a prophecy that helps fulfil it. Modern novels and films change the setting, voice, politics, psychology and meaning, but the underlying shape is often recognisable.

Familiarity is not necessarily a creative failure. A familiar structure can give an audience somewhere to stand while the creator changes what the story asks them to see.

Psychologists have been describing parts of creativity in similarly associative terms for decades. Sarnoff Mednick’s 1962 theory defined creative thinking as bringing associative elements into new combinations that are useful, with more remote combinations treated as more creative. The details of that model have since been challenged, but later research still distinguishes between generating remote associations and applying controlled processes to combine them under constraints. A 2018 study published in Brain examined the different brain networks involved in those two activities.

The important point is not that creativity has been scientifically reduced to recombination. It has not. It is that recombination has long been recognised as part of human creative thought.

The creative act is not only producing possibilities. It is recognising that two things which were previously separate belong together, then shaping the connection until it becomes useful, affecting or true.

What AI changes is the reachable territory

Human thinking has limits that have nothing to do with intelligence.

There is only so much information a person can read, remember and compare at once. Time forces us into familiar routes. We search for what we already know how to describe. We rely on the sources we remember. The first convincing explanation can become the frame through which everything else is interpreted.

This is where AI can change creative work in a more interesting way than simply generating the first draft.

It can make more territory reachable.

An AI system can compare a research report with interview transcripts, previous articles, campaign performance and an observation recorded months earlier. It can suggest that a problem in accounting resembles one in another profession. It can retrieve an abandoned idea when new evidence makes it relevant. It can show that two apparently different arguments depend on the same assumption. It can generate counterarguments, adjacent examples and alternative structures quickly enough that a person can inspect several routes before committing to one.

I have written that AI cannot supply my triggers. The initial observation still tends to come from something I noticed: a line in a film, an awkward piece of work, a customer comment, an argument that seems too neat. The system cannot decide what I encounter or why it catches my attention.

But that is only the first half of creative thought.

Once the observation enters the system, AI can help connect it to a much larger body of evidence than I could keep active in my head. It can give an old thought another chance to meet the piece of information that makes it useful.

That is why persistent context matters. In a blank chat, the model is largely working from the average of its training and whatever fits inside the immediate prompt. Inside an accumulated body of work, it can compare the new observation with specific evidence, previous arguments, decisions and contradictions. As I argued in Context is capital, old thinking can participate in new thinking.

This is not AI replacing creativity. It is AI changing the practical range across which a person can exercise it.

Hidden Hours and the cost of connection

Hidden Hours is the clearest example from my own work because the final narrative was not sitting inside any single document.

The research showed that accountants and bookkeepers were taking on substantial work outside agreed scope. Eighty-one per cent regularly did work beyond the brief. Seventy per cent said their fees did not reflect the full support they delivered. Beyond-scope work was linked to stress, reduced effectiveness and pressure on the time available for core work.

Those findings were valuable on their own. They could have produced a research report, a set of statistics and several useful articles.

The more important connection came from placing that evidence beside other themes that would usually have been planned separately: Making Tax Digital, AI, growth, embedded services and rising client expectations.

Once those signals were considered together, they began to describe the same underlying movement. Faster systems were not simply removing work. They were relocating pressure into judgement, verification, follow-up, responsiveness, coordination and client support — the parts of professional work that are hardest to see, scope and price.

That led to the organising thought behind the wider Winning in Small work: speed does not remove pressure; it relocates it.

AI did not invent that argument and it could not decide whether the argument was important. The useful role was more practical. It made it possible to compare enough research, planning, customer context and adjacent themes for the pattern to become visible without turning the exercise into weeks of manual processing.

I could probably have reached the same connection without AI. I am much less certain I could have reached it within the time a real organisation would allow.

That distinction matters.

We sometimes talk about AI as though the choice is between a human having the idea independently and a machine having it instead. In practice, the choice is often between a person exploring a wider set of connections with AI and the same person taking the most obvious available route because the deeper investigation costs too much.

This is central to Editorial Intelligence. The method begins with what an organisation knows, connects evidence that normally lives in separate projects and applies editorial judgement to decide what the combined evidence means. AI did not make that upstream work valuable. It made more of it affordable.

Hidden Hours also shows why the human role cannot be reduced to approving machine output. The important decision was not merely to accept a suggested theme. It was to recognise that several pressure systems belonged inside one defensible narrative, check that the evidence supported the connection, and reject interpretations that were fluent but too broad.

The machine expanded the possible routes. Editorial judgement gave the work direction.

The same machine can narrow thought

None of this means AI is naturally a machine for originality.

It is often a machine for convergence.

A 2024 experiment published in Science Advances gave some participants access to ideas from a generative AI system before they wrote short stories. The AI-assisted stories were judged more creative, better written and more enjoyable, with the greatest gains among people who had initially scored as less creative. But the stories also became more similar to each other.

That is the tension in unusually clear form. AI can improve an individual’s output while reducing the diversity of the group.

The same thing can happen in editorial work. If everybody asks similar models for an article on the same topic, accepts the first plausible structure and publishes the result, the individual piece may be competent while the wider information environment becomes narrower.

But it is important to remember that this failure did not begin with generative AI.

The SEO content system had already trained people to converge. Start with a keyword. Read the pages that already rank. Reproduce the subjects they cover. Use a similar structure, answer the same questions and make the article slightly longer. Much of the resulting material was written manually, slowly and professionally. It was still derivative.

I do not get to criticise that system from a clean distance. I worked inside the content industry it created. Writers were often asked to find an original angle while working from briefs, search results, deadlines and budgets designed to reward familiarity. The most practical route was to process what already existed and rearrange it into another acceptable version.

AI can industrialise that failure. It can create more derivative material, faster, at a scale human production could never reach.

It can also be used to interrogate the failure.

Instead of asking for the article, ask what assumptions every existing article shares. Ask which evidence contradicts the accepted frame. Compare the problem with an adjacent discipline. Retrieve what customers say that the market language leaves out. Ask for several incompatible interpretations, then test each against the evidence. Use one model’s answer as material for another model to challenge.

The model is still working from inherited material. The difference is whether the workflow asks it to collapse that material into an answer or open it into routes.

Non-linear thinking still needs judgement

There is a danger in romanticising non-linear thinking too.

More connections do not automatically produce a better idea. A remote association can be surprising and meaningless. A model can produce a clever analogy that collapses under scrutiny. It can connect two weak claims and make the result sound stronger than either one deserves. It can send an investigation into endless adjacent territory until the original question disappears.

I have done this often enough to know that AI can make thinking more expansive and more distracted at the same time.

Non-linear should not mean directionless.

The work still needs a centre: a question worth answering, evidence capable of changing the answer and a person willing to decide when the connection is strong enough to build. That person also has to decide what has merely been borrowed, what should be discarded, what deserves to be kept and what only they could have supplied.

This is where editorial judgement becomes more important rather than less.

When possible routes are scarce, producing another possibility feels creative. When possibilities are abundant, the creative work moves towards selection, combination and commitment.

Which route reveals something the evidence can support? Which familiar template helps the audience understand rather than merely reassuring them? Which contradiction should remain unresolved? Which connection is genuinely new in this context? Which idea would not exist without the organisation’s research, customers or experience?

AI cannot answer those questions in any final sense because it does not carry the consequences of the answer. It does not know which claim an organisation can legitimately make. It does not know which customer experience deserves more weight. It does not know when a surprising connection matters to the people involved rather than merely sounding intelligent.

The human role is not to contribute some mystical purity that the machine cannot contaminate. It is to provide attention, experience, evidence, intention and responsibility — then exercise judgement over a much larger field of possibility.

So I do not think the right question is whether AI itself is creative.

The more useful question is what kind of thinking our use of it encourages.

Used as a production engine, it can take us to the conventional answer faster. Used as an iterative synthesis partner, it can let us explore more of the territory before deciding what we think.

The originality still does not belong to the machine alone. It does not belong to an isolated human mind alone either.

It emerges from the relationship between what was noticed, what could be retrieved, what became connected and what somebody finally chose to make from it.